Abstract: In the existing fault diagnosis method of mine belt conveyor roller, the roller signal is decomposed and converted to the frequency domain. The fault diagnosis is carried out by extracting characteristics from the frequency domain. The common signal decomposition methods include wavelet decomposition and empirical mode decomposition. The methods have the problems of difficult selection of wavelet basis, frequency spectrum aliasing and endpoint effect, resulting in low fault diagnosis accuracy rate. In view of the above problems, an audio fault diagnosis method of mine belt conveyor roller mine based on variational modal decomposition (VMD)-BP neural network is proposed. Firstly, the audio signal of the roller along the mine belt conveyor is collected by the audio sensor. The audio signal is preprocessed to suppress the noise signal in the audio information. Secondly, VMD is used to decompose the audio signal into different IMF (intrinsic mode function) components according to the center frequency. The method extracts characteristic values of the kurtosis, gravity frequency, frequency standard deviation of each IMF component. Finally, the characteristic values are input into the trained BP neural network. According to the difference in IMF component characteristic values, it is possible to diagnose the mine belt conveyor roller fault through audio, and determine the position of the faulty roller according to the sensor number corresponding to the audio signal. The audio information of the roller of the belt conveyor collected in a coal mine is used to analyze and verify the audio fault diagnosis method of mine belt conveyor roller based on VMD-BP neural network. The results show that the method can avoid spectrum aliasing and endpoint effect in the decomposition process when decomposing and extracting audio signal characteristics. The overall fault diagnosis accuracy rate reaches 96.15%. Compared with the fault diagnosis method based on BP neural network and the fault diagnosis method based on wavelet decomposition and BP neural network, the proposed method has improved the fault diagnosis accuracy rate by 26.92% and 15.38% respectively. The false detection rate has also been significantly reduced.